CATTest - Bug Discovery R-score: leaderboard
Fine-grained R-score for how many annotated ground-truth bugs a VLM testing agent uncovers while writing Playwright code within CATJudge to explore 102 AI-generated web applications; per-project scores are summed and normalized (CATTest, arXiv 2609.00081v1, Table 1). 0-100; higher is better.
Metric: R-score (0-100; recall-oriented bug discovery). Source: arxiv.org. Saturation forecast: Around July 2027. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 42.57 | #51 |
| 2 | Claude Opus 4.6 | 41.55 | #67 |
| 3 | GPT-5.5 | 39.41 | #33 |
| 4 | Claude Sonnet 4.6 | 37.68 | #94 |
| 5 | Gemini 3.1 Pro (Preview) | 36.49 | #64 |
| 6 | Qwen 3.6 27B | 33.56 | #182 |
| 7 | Kimi K2.6 | 30.78 | #113 |
| 8 | Qwen 3.5 Plus | 24.71 | #133 |
| 9 | GLM-5V Turbo | 24.67 | #172 |
| 10 | Gemini 3 Flash | 24.22 | #103 |
| 11 | Qwen 3.5 27B | 23.24 | #192 |
| 12 | Seed 2.0 Pro | 22.79 | #122 |
| 13 | Seed 2.0 Lite | 21.98 | #161 |
| 14 | Gemini 3.1 Flash Lite | 20.92 | #200 |
| 15 | GPT-5.4 Mini | 19.07 | #228 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=cattest-bug-discovery-r-score · How It Works · Data refreshed daily, snapshot 2026-11-05.